Está en la página 1de 9

Journal of Materials Processing Technology 142 (2003) 2028

Delta ferrite prediction in stainless steel welds using neural network analysis and comparison with other prediction methods
M. Vasudevan a, , A.K. Bhaduri a , Baldev Raj a , K. Prasad Rao b
a

Metallurgy and Materials Group, Indira Gandhi Centre for Atomic Research, Kalpakkam, India b Department of Metallurgy, Indian Institute of Technology, Chennai, India Received 2 May 2002; received in revised form 11 December 2002; accepted 17 February 2003

Abstract The ability to predict the delta ferrite content in stainless steel welds is important for many reasons. Depending on the service requirement, manufacturers and consumers often specify delta ferrite content as an alloy specication to ensure that weld contains a desired minimum or maximum ferrite level. Recent research activities have been focused on studying the effect of various alloying elements on the delta ferrite content and controlling delta ferrite content by modifying the weld metal compositions. Over the years, a number of methods including constitution diagrams, Function Fit model, Feed-forward Back-propagation neural network model have been put forward for predicting the delta ferrite content in stainless steel welds. Among all the methods, neural network method was reported to be more accurate compared to other methods. A potential risk associated with neural network analysis is over-tting of the training data. To avoid over-tting, Mackay has developed a Bayesian framework to control the complexity of the neural network. Main advantages of this method are that it provides meaningful error-bars for the model predictions and also it is possible to identify automatically the input variables which are important in the non-linear regression. In the present work, Bayesian neural network (BNN) model for prediction of delta ferrite content in stainless steel weld has been developed. The effect of varying concentration of the elements on the delta ferrite content has been quantied for Type 309 austenitic stainless steel and the duplex stainless steel alloy 2205. The BNN model is found to be more accurate compared to that of the other methods for predicting delta ferrite content in stainless steel welds. 2003 Elsevier Science B.V. All rights reserved.
Keywords: Neural network analysis; Delta ferrite content; Austenitic stainless steel; Duplex stainless steel

1. Introduction The ability to estimate the delta ferrite content accurately has proven very useful in predicting the various properties of austenitic SS welds. A minimum delta ferrite content is necessary to ensure hot cracking resistance in these welds [1,2], while an upper limit on the delta ferrite content determines the propensity to embrittlement due to secondary phases, e.g. sigma phase, etc., formed during elevated temperature service [3]. At cryogenic temperatures, the toughness of the austenitic SS welds is strongly inuenced by the delta ferrite content [4]. In duplex stainless steel weld metals, a lower ferrite limit is specied for stress corrosion cracking resistance while the upper limit is specied to ensure adequate ductility and toughness [5]. Hence, depending on the service requirement a lower limit and/or an upper limit on delta ferrite content is generally specied. During the selec-

Corresponding author. Tel.: +91-4114-80232; fax: +91-4114-40381. E-mail address: dev@igcar.ernet.in (M. Vasudevan).

tion of ller metal composition, the most accurate diagram to date WRC-1992 is used generally to estimate the -ferrite content [6]. The Creq and Nieq formulae used for generating the WRC-1992 constitution diagram is given by Creq = Cr + Mo + 0.7Nb and Nieq = Ni + 35C + 20N + 0.25Cu. The limitation of these equations is that values of the coefcients for the different elements remain unchanged irrespective of the change in the base composition of the weld. However, the relative inuence of each alloying addition given by the elemental coefcients in the Creq and Nieq expressions is likely to change over the full composition range. Furthermore, these expressions ignore the interaction between the elements. Also, there are a number of alloying elements that have not been considered in the WRC-1992 diagram. Elements like Si, Ti, W have not been given due to considerations, though they are known to inuence the delta ferrite content. Hence, the delta ferrite content estimated using the WRC-1992 diagram would always be less accurate and may never be close to the actual measured value. In the Function Fit model [7] for estimating ferrite, the difference in free energy between the ferrite and the austenite was calculated

0924-0136/$ see front matter 2003 Elsevier Science B.V. All rights reserved. doi:10.1016/S0924-0136(03)00430-8

M. Vasudevan et al. / Journal of Materials Processing Technology 142 (2003) 2028

21

as a function of composition and this was related to ferrite number (FN). The equation used in this model to determine FN is given below: FN = A[1 + exp(B + C G)]1 (1)

Table 1 Range, mean and standard deviations of the composition variables (input) and the FN (output) Elements C Mn Si Cr Ni Mo N Nb Ti Cu V Co Fe FN Minimum value 0.0000 0.3500 0.0300 1.0500 4.6100 0.0100 0.0100 0.0000 0.0000 0.0000 0.0000 0.0000 45.5990 0.0000 Maximum value 0.2000 12.6700 6.4600 32.0000 33.5000 10.7000 2.1300 0.8800 0.3300 6.1800 0.2300 0.6900 72.5150 98.0000 Mean 0.0401 1.8805 0.5255 20.5135 11.3062 1.4186 0.0889 0.0287 0.0200 0.1405 0.0373 0.0300 63.9412 12.0358 Standard deviation 0.0219 1.7897 0.3477 2.7615 2.5628 1.6363 0.1415 0.0982 0.0284 0.4368 0.0404 0.0463 4.3320 17.3052

where A, B and C are the constants. The advantages of this semi-empirical model over the WRC-1992 diagram include its considering effect of other alloying elements and the ease of extrapolation to higher Creq and Nieq values. This Function Fit method can be used for a wide range of weld metal compositions and owing to the analytical form of this model, the FN can be quantied easily. However, the accuracy of this method is not greater than the WRC-1992 diagram. Vitek et al. [8,9] sought to overcome the major limitation of the constitution diagram and the Function Fit method of not taking into account the elemental interactions, by using neural networks for predicting ferrite in SS welds. The improvement in accuracy in predicting the delta ferrite content by using neural networks, involving a feed-forward network with a back-propagation optimization scheme, has been clearly brought in their study. The effect of various elements on the delta ferrite content for a few base compositions was examined by calculating the FN as a function of composition. However, it was not possible in their analysis to directly interpret the elemental contributions to the nal FN. The prediction and measurement of ferrite in SS welds remains of scientic interest due to limitations in all the current methods, and newer methods and constitution diagrams are continuously being proposed to predict the delta ferrite content for a wider range of SS types. It was in this context that the development of a more accurate neural network based predictive tool for estimating the effect of various alloying elements on the delta ferrite content for different SS welds was taken up in this work. A potential risk associated with neural network analysis is over-tting of the training data. To avoid over-tting, Mackay [10] developed a Bayesian framework to control the complexity of the neural network, with its main advantages being that it provides meaningful error-bars for predictions and also enables identifying the input variables that are important in the non-linear regression. Hence, in the present study, Bayesian neural network (BNN) analysis was applied to develop a generalized model for FN prediction in stainless steel welds and the effect of variations in the concentration of the elements on the FN for 309 stainless steel and duplex stainless steel base compositions were also quantied.

generating the WRC-1992 diagram was used [11]. For the datasets in which the composition values for elements such as Nb, Ti, V, Cu and Co were not available, their values were assumed zero. Table 1 shows the range, mean and standard deviation of the each composition variable (input) and the FN (output). This simply gives the idea of the range covered and cannot be used to dene the range of applicability of the neural network model as the input variables are expected to interact in neural network analysis. In BNN analysis, size of the error-bars dene the range of useful applicability of the trained network.

3. BNN analysis The networks employed consist of 13 input nodes xi representing the 13 composition variables, a number of hidden nodes hi and one output y. The schematic structure of the network is shown in Fig. 1. The single output represents the FN. Both the input and output variables were normalized within the range 0.5 as follows: xN = x xmin 0.5 xmax xmin (2)

2. Database As the aim of the present work was to model for the FN as a function of chemical composition, the database of 924 datasets for shielded metal arc (SMA) weld compositions and delta ferrite contents, representing the common 300-series SS weld compositions (viz., 308, 308L, 309, 309L, 316, 316L, etc.) and duplex stainless steels used for

where xN is the normalized value of x, which has maximum and minimum values given by xmax and xmin . Eighty different neural network models were created using the datasets, with the number of the hidden units varying from 1 to 16 and with ve different sets of random seeds used to initiate the network for a given number of hidden units. All these models were trained on a training dataset that consisted of a random selection of half the datasets (i.e. 462 datasets), while the remaining half formed the test dataset that was used to examine how the model generalizes with unseen data. For calculating the outputs from the inputs, the linear functions of the inputs xj are multiplied by the weights wij and operated by the following hyperbolic tangent transfer

22

M. Vasudevan et al. / Journal of Materials Processing Technology 142 (2003) 2028

Fig. 1. Schematic diagram of the network structure showing the input nodes, hidden units and the output node.

function so that each input contributes to every hidden unit, where N is the total number of input variables: hi = tanh
N j (1) (1) wij xj + i

(3)

Here the bias is designated as and is analogous to the constant in linear regression. The transfer from the hidden units to the output is linear, and is given by:
N

for making a wild prediction is reduced if that prediction is accompanied by an appropriately large error-bar, with a larger value of the LPE implying a better model. Further, by this method it is also possible to automatically identify the input variables that are signicant in inuencing the output variable, as the input variables that are less signicant are down-weighted in the regression analysis. 3.1. Over-tting problem In BNN analysis, two solutions are implemented which contribute to avoid over-tting. The rst is contained in the algorithm due to MacKay [12]: the complexity parameters and are inferred from the data, therefore allowing automatic control of the model complexity. The second resides in the training method. The database is equally divided into a training set and a testing set. To build a model, about 80 networks are trained with different number of hidden units and seeds, using the training set; they are then used to make predictions on the unseen testing set and are ranked by LPE. 3.2. Committee model The networks with different number of hidden units will give different predictions. But predictions will also depend on the initial guess made for the probability distribution of the weights (the prior). Optimum predictions are often made using more than one model, by building a committee. The prediction y of a committee of networks is the average prediction of its members, and the associated error-bar is calculated according to Eq. (6): y = 1 L 1 L y(l)
l (l) y +
2

y=
i

wi hi + (2)

(2 )

(4)

The output y is therefore a non-linear function of xj , with the function usually selected for exibility being the hyperbolic tangent. Thus, the network is completely described if the number of input nodes, output nodes and the hidden units are known along with all the weights wij and biases i . These weights wij are determined by training the network and involves minimization of a regularized sum of squared errors. The BNN analysis of Mackay [10] allows the calculation of error-bars with two componentsone representing the perceived level of noise (v ) in the output and the other indicating the uncertainty in the data tting. This latter component of the error-bars emanating from the Bayesian framework allows the relative probabilities of the models with different complexity to be assessed. This enables estimation of quantitative error-bars, which vary with the position in the input space depending on the uncertainty in tting the function in that space. Hence, instead of calculating a unique set of weights, a probability distribution of weights is used to dene the uncertainty in tting. Therefore, these error-bars become large when data are sparse or locally noisy. In this context, a very useful measure of the error is the logarithm of the predictive error (LPE) given by the following: LPE =
n

2 =

1 2

(t (n)

y(n) )2
(n)2

n 1/2 + log(2y )

1 L

(y(l) y )2
l

(6)

(5)

where t is the target for the set of inputs x, while y the corresponding network output. y is related to the uncertainty of tting for the set of inputs x. By using LPE, the penalty

where L is the number of networks in a committee. Note that we now consider the predictions for a given single set of inputs and that exponent l refers to the model used to produce the corresponding prediction y(l ) . In practice, an increasing number of networks are included in a committee and their

M. Vasudevan et al. / Journal of Materials Processing Technology 142 (2003) 2028

23

performances are compared on the testing set. Most often, the error is minimum when the committee contains more than one model. The selected models are then retrained on the full database.

4. Results and discussions The characteristics of the BNN model on FN is discussed in detail elsewhere [13]. The comparison between the predicted and measured FN values for the committee of models (38 models in the committee) is shown in Fig. 2 for the complete dataset. There was excellent agreement between the measured and the predicted FN values. The correlation coefcient was determined to be 0.98025. 4.1. Signicance of the individual elements on the FN Fig. 3 indicates the signicance w of each of the input variables as perceived by the rst ve neural network models in the committee. The w value represents the extent to which a particular input explains the variation in the output, as for a partial correlation coefcient in linear regres-

sion analysis. It is observed from Fig. 3 that the elements Mn and Nb are not signicant in inuencing the FN. The observation of Mn not having a signicant inuence on the FN for the 300-series austenitic SS is in agreement with the reported results that the variation in Mn from 1 to 12 wt.% had almost no effect on the as-deposited FN [14]. However, the element Nb, which was found in the present study to have an insignicant effect on the FN is included in the Creq formula used in the WRC-1992 diagram. As expected, Cr and Ni were found to be the main elements inuencing the FN. As per the present model, the elements that inuence the FN in order of signicance are: Mo > N > V > Ti > Cu > Co > Si > C > Fe. However, some of these elements namely V, Ti, Co and Si are not included in the Nieq and Creq formulae used in the WRC-1992 diagram. 4.2. Comparison of accuracy of present model with other existing methods Analysis of the error distributions (measured FNpredicted FN) for the present BNN model shows that the absolute error was <2.5 for most of the datasets used in the training, while for the FNN-1999 model the absolute error was <3 for about 80% of the dataset used in training. It is important to note here that in the present BNN model, the entire datasets were used for retraining the committee of models, while in the FNN-1999 model only 90% of the datasets were used for training. Further, the error distributions for the present BNN model is symmetrical about zero (Fig. 4) implying good tting of the model to the datasets. Also, the tail of the error distributions is less compared to the other methods. Table 2 shows the comparisons of the errors for the BNN model and the FNN-1999 model. Comparison of the quantied error distributions with those of the FNN-1999 model shows that the present BNN model is superior. It has been reported that the FNN-1999 model is more accurate compared to the WRC-1992 diagram and Function Fit model. Hence, the root mean square (RMS) error between the measured and predicted FN values for the present BNN model and the other three methods were compared and it was found that the present BNN model

Fig. 2. Comparison between the predicted and measured FN values for the entire dataset using the committee of models.

Fig. 3. Perceived signicance w values of the rst ve FN models for each input.

Fig. 4. Error distributions (experimental FNpredicted FN) for the complete dataset for BNN model.

24

M. Vasudevan et al. / Journal of Materials Processing Technology 142 (2003) 2028

has the lowest error among the four methods, BNN model showing an improvement of 43% over the FNN-1999 model and about 65% over the WRC-1992 diagram and Function Fit model. Table 3 shows the RMS error values for all the four methods. As the RMS error values represent the quan-

titative measure of the degree of t of the various models to the datasets on which they were trained, this comparison clearly establishes that among the available methods the present BNN model is the most accurate model for prediction of FN in austenitic SS welds.

Fig. 5. Predicted FN vs concentration of the elements for 309 austenitic stainless steels weld. The plot shows the variation in the FN when one of the element is varied and all other concentration are held constant at the 309 SS composition except Fe, which is adjusted to compensate for the varying element concentration.

M. Vasudevan et al. / Journal of Materials Processing Technology 142 (2003) 2028

25

Fig. 5. (Continued ).

Table 2 Comparison of the errors (experimental FNpredicted FN) for the BNN model and the FNN-1999 model (training database) BNN model Absolute error 1.5 2.5 3.5 4.5 5.5 5.5 9.5 Number of points 684 820 864 888 900 20 4 Total no. of points (%) 74.0 88.7 93.5 96.1 97.4 2.16 0.4 FNN-1999 model Number of points 621 764 826 32 Total no. of points (%) 64.6 79.5 86 3.3

4.3. Effect of compositional variations on the FN The severe limitation of the WRC-1992 diagram is that the coefcients in the terms for Creg and Nieg formulas are constant and hence the inuence of an individual element on FN is same irrespective of the change in the base composition. As neural networks can take into account the interaction between the input variables on their inuence over the output variable, the interaction between the different elements on their inuence over the FN is quantied for stainless steel welds using the BNN analysis. The results for 308, 308L, 316, 316LN have already been presented elsewhere [13,15]. In the present work, the effect of variations in the concentration of the elements on FN have been quantied for 309 stainless steel and duplex stainless steel welds. This was done with two starting base compositions and then alTable 3 Comparison of the RMS errors for complete training database for different FN prediction methods Prediction method BNN model [12] FNN-1999 (back-propagation neural network) model [8] WRC-1992 [6] Function Fit model [7] RMS error 1.99 3.5 5.8 5.6

lowing each element to vary over a limited range adjusting Fe concentration accordingly but holding all other element concentrations constant. Table 4 shows the base compositions of the 309 SS and duplex stainless steel welds used in the present study. 4.3.1. The 309 stainless steel weld The variation in the predicted FN as a function of the variation in the concentration of the elements is found to be non-linear (Fig. 5). The FN is found to decrease with increasing concentration of the elements C, N and Ni. These elements acts as austenite stabilizers. The FN is found to increase with increasing concentration of the elements Cr, Si and V and these elements are called ferrite stabilizers. The above observations are in agreement with the literature. The elements Mn, Mo, Nb, Ti, Cu and Co do not inuence
Table 4 Chemical composition of the 309 SS weld and duplex stainless steel weld Material 309 DSS 2205 N 0.09 0.12 C 0.055 0.035 Nb 0.02 0.04 Mn 1.5 1.4 Ti 0.02 0.08 Si 0.4 0.4 Cu 0.03 0.12 Cr 23 22.25 V 0.08 0.10 Ni 12.5 9.25 Co 0.03 0.09 Mo 0.02 4.0 Fe 62.255 62.115

26

M. Vasudevan et al. / Journal of Materials Processing Technology 142 (2003) 2028

the FN value for 309 base composition used. However, in the WRC-1992 diagram the composition of the element Cu is taken in to account for calculating the Nieq and the composition of Mo and Nb are included for calculating

the Creq . Thus, estimation of delta ferrite content by using the WRC-1992 diagram will always be less accurate. The BNN model generated by us is more accurate compared to the WRC-1992 diagram which was generated based on the

Fig. 6. Predicted FN vs concentration of the elements for duplex stainless steel weld. The plot shows the variation in the FN when one of the element is varied and all other concentration are held constant at the duplex stainless steel composition except Fe, which is adjusted to compensate for the varying element concentration.

M. Vasudevan et al. / Journal of Materials Processing Technology 142 (2003) 2028

27

Fig. 6. (Continued ).

linear regression analysis. Hence, the trends of the inuence of concentration of the elements on FN predicted by the model is more useful in controlling FN through compositional modications in this type of steel (Fig. 5). 4.3.2. Duplex stainless steel (alloy 2205) weld The variation in the FN with variation in the concentration of the elements is found to be non-linear (Fig. 6). The increase in the concentration of the elements C, N, Mn and Ni is found to decrease the FN. However, the effect of Mn is not as signicant as the other austenite stabilizers. The increase in the concentration of the elements Cr, Si, Mo, V and Co is found to increase the FN for the duplex stainless steel welds. The effect of vanadium is not as signicant as the other ferrite stabilizers. The elements Cu, Nb and Ti are found not to inuence the FN for duplex stainless steel welds. However, the elements Cu and Nb are included in the WRC-1992 diagram in calculating the Nieq and Creq , respectively. The trends identied by this analysis of the inuence of concentration of the elements on the FN is very useful in controlling the FN by adjusting weld metal composition in duplex stainless steel welds. Hence, depending on the base composition, the inuence of individual elements on the FN is different. However, the WRC-1992 diagram uses the same equation for all the stainless steel welds and is the severe limitation of the diagram (Fig. 6).

3. The effect of variation in the concentration of the elements on the FN have been quantied for 309 and duplex stainless steel welds. Neural network analysis has shown that there is a change in the role of elements when the base composition is changed. 4. Cobalt is found to be ferrite stabilizer for the duplex stainless steel welds and is found not to inuence the FN for the austenitic stainless steel welds.

References
[1] C.D. Lundin, C.P.D. Chou, Hot cracking susceptibility of austenitic stainless steel weld metals, WRC Bull. 289 (1983) 180. [2] C.D. Lundin, W.T. Delong, D.F. Spond, Ferrite-ssuring relationships in austenitic stainless steel, Weld Met. 54 (8) (1975) 241s246s. [3] J.M. Vitek, S.A. David, The sigma phase transformation in austenitic stainless steels, Weld. J. 65 (4) (1986) 106s111s. [4] E.R. Szumachowski, H.F. Reid, Cryogenic toughness of SMA austenitic stainless steel weld metals, Weld. J. 57 (11) (1978) 325s 333s. [5] D.J. Kotecki, Ferrite control in duplex stainless steel weld metal, Weld. J. 65 (10) (1986) 273s278s. [6] D.J. Kotecki, D.T.A. Siewert, WRC-92 constitution diagram for stainless steel weld metals: a modication of the WRC-1988 diagram, Weld. J. 71 (5) (1992) 171s178s. [7] S.S. Babu, J.M. Vitek, Y.S. Iskander, S.A. David, New model for prediction of ferrite number in stainless steel welds, Sci. Technol. Weld. 2 (6) (1997) 279285. [8] J.M. Vitek, Y.S. Iskander, E.M. Oblow, Improved ferrite number prediction in stainless steel arc welds using articial neural networks. Part 1. Neural network development, Weld. J. 79 (2) (2000) 33 40. [9] J.M. Vitek, Y.S. Iskander, E.M. Oblow, Improved ferrite number prediction in stainless steel arc welds using articial neural networks. Part 2. Neural network development, Weld. J. 79 (2) (2000) 41 50. [10] D.J.C. Mackay, Bayesian non-linear modeling with neural networks, in: H. Cerjack (Ed.), Mathematical Modeling of Weld Phenomena, vol. 3, The Institute of Materials, London, 1997, pp. 359389. [11] C.N. McCowan, T.A. Siewert, D.L. Olson, Stainless steel weld metal: prediction of ferrite content, WRC Bull. 342 (1989) 136.

5. Conclusions 1. The generalized model for predicting the FN in stainless steel welds using BNN analysis has been developed. The accuracy of the BNN model in predicting FN is superior compared to the existing FN prediction methods. 2. Signicance of the individual elements on FN has been quantied. Neural network analysis has shown that elements like manganese and niobium are insignicant in inuencing the FN in stainless steel welds.

28

M. Vasudevan et al. / Journal of Materials Processing Technology 142 (2003) 2028 [14] E.R. Szumachowski, D.J. Kotechi, Effect of manganese on stainless steel weld metal ferrite, Weld. J. 63 (5) (1984) 156s161s. [15] M. Vasudevan, A.K. Bhaduri, B. Raj, K. Prasad Rao, Bayesian neural network analysis of the compositional variations on the ferrite number in 316LN austenitic stainless steel welds, Trans. Ind. Ins. Met. 55 (5) (2002) 389396.

[12] D.J.C. MacKay, A practical Bayesian framework for backpropagation networks, Neural Comput. 3 (1992) 448472. [13] M. Vasudevan, M. Murugananth, A.K. Bhaduri, Application of Bayesian neural network for modeling and prediction of FN in austenitic stainless steel welds, in: H. Cerjak, H.K.D.H. Bhadeshia (Eds.), Mathematical Modelling of Weld PhenomenaVI, Institute of Materials, 2002, pp. 10791099.

También podría gustarte